Register a spatial cache. The spatial cache can be any thing you want to cache. The registery and retrieval of the cache is b

Register a spatial cache. The spatial cache can be any thing you want to cache. The registery and retrieval of the cache is b

r"""Sample spatio-temporal features `sample_features5d(input, coords)` works in the same way as :func:`sample_features4d` but for spatio-temp

logits: Tensor of arbitrary shape targets: same shape as logits balance_weight: scaling the loss reduction: 'mean', 'sum', or 'none'

Unproject a depth map to a point cloud in COLMAP convention. Args: metric_depth: (B, H, W) depth map, meters. intrinsics:

UV with left-top corner as (-width / diagonal, -height / diagonal) and right-bottom corner as (width / diagonal, height / diagonal)

Cat x with a positional encoding of x with scales 2^[min_deg, max_deg-1]. Instead of computing [sin(x), cos(x)], we use the trig identity cos

@pts: (N,3 or 2) will homogeneliaze the last dimension

This function will generate tracks by camera transform Args: queries: B T N 4 c2w_traj: B T 4 4

Extract some coefficients at specified timesteps, then reshape to [batch_size, 1, 1, 1, 1, ...] for broadcasting purposes.

Initialize the image conditioning model.

If trunc is None, solve `min sum_i w_i * |a * x_i - y_i|`, otherwise solve `min sum_i min(trunc, w_i * |a * x_i - y_i|)`. w_i must be >=

If trunc is None, solve `min sum_i w_i * |a * x_i - y_i|`, otherwise solve `min sum_i min(trunc, w_i * |a * x_i - y_i|)`. w_i must be >=

This function will generate tracks by camera transform Args: queries: B T N 3 scale_est: 1 1 shi

Link the output of each node to the input of the next node.

@depths: torch tensor (B,H,W) @Ks: torch tensor (B,3,3)

embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)

@type: mask_visib (only visible part) / mask (projected mask from whole model)

Annearl truncation over training

@imgs: (B,H,W,C) np array @nrow: num of images per row

Sample from a model.

Zero out the parameters of a module and return it.

Init. Args: scale_factor (float): scaling mode (str): interpolation mode

Compute the image features on a given frame.

Predict masks for the given input prompts, using the currently set image. Input prompts are batched torch tensors and are expected to

Delete file or directory after specified delay (default 10 minutes)

Load a pretrained model.

Get SAM predictor with option to use HuggingFace version Args: model_type: Model type ('vit_b', 'vit_l', 'vit_h') device: Device t

This function is used to get the points on the grid args: H: the height of the grid. W: the width of the grid. T: the

Initialize and return the tracker predictor and visualizer Args: output_dir: Directory to save visualization results vo_point

Initialize a inference state.

Faster `np.linalg.norm(x, axis=-1)` for 3D vectors

Return a normalization layer.

Args: pose_encoding: A tensor of shape `BxNxC`, containing a batch of `BxN` `C`-dimensional pose encodings.

Preprocess the input image using BiRefNet for background removal. Includes padding to maintain aspect ratio when resizing to 518x518.

Recover the depth map and FoV from a point map with unknown z shift and focal. Note that it assumes: - the optical center is at the cent

Discard class token and reshape 1D feature map to a 2D grid.

Apply scaled dot product attention. Args: qkv (torch.Tensor): A [N, L, 3, H, C] tensor containing Qs, Ks, and Vs.

Scatter the minimum value along the given dimension of `input` into `src` at the indices specified in `index`.